Action-Oriented Learning Design for AI Hardware Courses

Practice Paper Recommended citation: Gursch, A. M., Hasse, L., Stober, D., Trinitis, C., & Lucke, U. (2025). Action-Oriented Learning Design for AI Hardware Courses. In Kangaslampi, R., Langie, G., Järvinen, H.-M., & Nagy, B. (Eds.), SEFI 53rd Annual Conference. European Society for Engineering Education (SEFI), Tampere, Finland. DOI: 10.5281/zenodo.17631822. This Conference Paper is brought to you for open access by the 53rd Annual Conference of the European Society for Engineering Education (SEFI) at Tampere University in Tampere, Finland. This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License. 1Corresponding Author A.-M. Gursch [email protected] ACTION-ORIENTED LEARNING DESIGN FOR AI HARDWARE COURSES A.-M. Gursch a,1, L. Hasse b, D. Stober c, C. Trinitis d, U. Lucke e aUniversity of Potsdam, Potsdam, Germany, 0009-0005-9696-1571 bUniversity of Potsdam, Potsdam, Germany, 0009-0007-1560-3559 cTechnical University of Munich, Munich, Germany, 0009-0001-5096-7063 dTechnical University of Munich, Heilbronn, Germany, 0000-0002-6750-3652 eUniversity of Potsdam, Potsdam, Germany, 0000-0003-4049-8088 Conference Key Areas: Digital tools and AI in engineering education, Curriculum development and emerging curriculum models in engineering Keywords: AI hardware, 4C/ID model, AI ethics, curriculum design, higher education ABSTRACT The growing significance of artificial intelligence (AI) has increased the prominence of AI topics in university curricula. However, these offerings are mostly limited to fundamentals, algorithms, and applications, with specific hardware considerations often neglected. This article aims to address this gap. It proposes a method for media-based instructional design of AI hardware courses. First, we present the action-oriented foundation for course design and learning. Next, we position relevant courses within the framework of the Dagstuhl Triangle. We then outline the systematic adoption of such courses in an online format, which supports crossinstitutional settings and facilitates the transfer of the results to other topics or institutions. Finally, we showcase two exemplary courses, sharing the experiences gained from their implementation. 1 INTRODUCTION The need for new AI hardware study programs arises from AI's disruptive potential in digitization. To date, hardware considerations in the AI sector have predominantly focused on performance aspects. However, from a legal and ethical perspective, dedicated AI hardware — in terms of sensor platforms and other embedded systems — is crucial to ensure data privacy by reducing dependency on cloud-based processing. This makes edge computing (Beck et al., 2014) essential for aligning data innovations with societal values. While industry leaders are increasingly integrating AI into their platforms, there is a noticeable lack of educational focus on dedicated AI hardware. At the same time, universities currently have limited resources to extensively cover this subject, which supports the idea of creating collaborative online courses. This approach, in turn, presents specific challenges for teaching this hands-on topic from a didactic perspective. We are developing an educational program focused on AI hardware that seeks to address this gap through a hybrid, cross-university approach, offering open courses, practical learning, and collaborative teamwork. The consortium, including the University of Potsdam (UP) and the Technical University of Munich (TUM) uniquely enables chip manufacturing in Germany at the Leibniz Institute IHP, allowing students hands-on experience with AI hardware. Regarding content, the design of an educational program dedicated to AI hardware can build on existing courses in the fields of AI, computer engineering, and various application domains. From a didactic perspective, the course design process benefits from a solid methodological foundation provided by CS education and didactics of media in education. 2 SUBJECT-SPECIFIC AND MEDIA-DIDACTIC FOUNDATIONS 2.1 Current Status of AI Hardware Topics in CS Education There is a growing trend toward acknowledging the significance of AI hardware and integrating related topics into computer science (CS) programs in higher education. However, the extent to which dedicated AI hardware is addressed in CS curricula varies widely across institutions based on their focus and priorities. Some universities enhance their CS programs with specialized courses on hardware accelerators, graphics, and tensor processing units designed for machine learning (Xiong et al., 2023). Additionally, AI hardware concepts are sometimes integrated into broader machine learning courses, discussing hardware requirements for training and deployment. In some cases, interdisciplinary collaborations offer projectbased learning (Kuo & Wu, 2023), fostering a comprehensive understanding of AI hardware. These curricula are still evolving. Current curriculum recommendations from scientific associations largely overlook AI hardware topics. The joint ACM and IEEE CS guidelines approach AI exclusively from an algorithmic perspective (Joint Task Force on Computing Curricula, 2013), while computer engineering recommendations define AI as a complementary field (Joint Task Force on Computing Curricula, 2020). Recommendations on Data Science mention hardware only in the context of cluster computing (Danyluk & Leidig, 2021). Similar gaps exist in national recommendations, including those from the German Society for Computer Science (Maehle et al., 2018; Zukunft, O., 2016). While the new EU AI Act obliges higher education institutions to train their students in AI competence (Schmermund, 2024), hardware topics remain unaddressed. Furthermore, AI hardware is not yet regarded as relevant in school curricula and CS teacher training. 2.2 Theoretical Foundations of CS Education In addition to the curriculum recommendations cited above, theoretical frameworks such as the Dagstuhl Triangle (Brinda et al., 2016) and its variation, the Frankfurt Triangle (Brinda et al., 2019), offer established models for designing and evaluating CS curricula. These frameworks provide a systematic approach to aligning courses with the requirements of scientific disciplines, the labor market, and societal needs. A key focus is to go beyond the technical specifics of IT systems to include typical application areas and the role of individuals as active, media-engaged subjects within society. Given the limited number of courses, programs and research groups dedicated to AI hardware, cross-institutional collaboration and, as a consequence, the use of digital learning environments play a significant role. 2.3 Instructional Design Based on 4C/ID The strong emphasis on practical skills in circuit design makes action-oriented instructional design models particularly advantageous. The Four Component Instructional Design (4C/ID) model (van Merriënboer et al., 2022), for example, which focuses on developing complex skills and competencies, allows for the application of theoretical knowledge in real-world contexts. This model has been successfully applied in various settings such as vocational training using virtual reality (Zender et al., 2020) and data science education (Linxen et al., 2023). Additionally, 4C/ID has played a crucial role in the design of professional development programs, e.g. for teaching differentiation skills to primary school teachers (Frerejean et al., 2021). Another noteworthy application focuses on improving the transfer of skills from classroom to clinical practice (Maggio et al., 2015). These examples demonstrate the model's adaptability and impact across domains. We chose the 4C/ID model for course design because it accommodates the diverse prior knowledge of students from various disciplines, supports individualized learning paths, and emphasizes real-world tasks, fostering action-oriented learning. Instructional design based on 4C/ID is thus more targeted than with other approaches (Gagné et al., 2005; Niegemann et al., 2008). In 4C/ID course design, the subject is divided into task classes, each encompassing four key components: Learning tasks are based on real-world tasks. They address the subject matter with progressively increasing difficulty and decreasing support, with a focus on the application of knowledge and skills and the higher levels of Bloom's taxonomy (Bloom, 1956). Supportive information offers guidance on non-routine aspects of learning tasks by providing cognitive strategies and mental models, and is continuously available. Procedural information supports learners in performing routine aspects of learning tasks and is presented during task execution in the form of rules and checklists. Part-task practice offers targeted repetition of specific routine aspects of learning tasks, particularly those that demand a high level of automaticity. Task classes progress from addressing the requirements of the application field to focusing on the specifics of hardware design. 3 DESIGNING AN AI HARDWARE PROGRAM In order to open this relevant topic to a great number of students despite limited teaching capacities and to consolidate competences, we created a cross-institutional program with blended-learning courses open to both undergraduate and graduate students majoring in computer science, computer engineering, and computational science at one of the participating universities. In the following, we outline the design process, starting with the positioning of relevant courses in the Dagstuhl triangle and ending with the description of a course template to support the implementation of the 4C/ID model in Moodle. 3.1 Balancing Perspectives within the Dagstuhl Triangle Having identified relevant courses from the domains of AI, computer engineering, and the geosciences (as an application domain), we began by determining the general orientation and positioning of individual courses through a comparison of the courses and learning objectives. Grounded in the Dagstuhl triangle, our analysis comprehensively assessed the technical, socio-cultural, and application perspectives through collaboration with the respective educators. The aim was to critically examine the reasoning behind the classifications, fostering discussions on both the legitimacy of the courses and their alignment with research and industry demands. The resulting framework (shown in Fig. 1) serves as a foundation for course design. It includes four lectures with accompanying exercises, two practical courses, and two seminars, all of them yielding 5 to 6 ECTS. All courses took place 2 to 3 times. Our analysis of the courses reveals a focus on the technological and application perspectives of AI hardware, marginalizing the societal-cultural dimension. Only the seminar "Ethics for Nerds" (no. 8 in Fig. 1) addresses this perspective, prompting questions about the distribution's appropriateness and the potential need for more courses in this area. Lecturers debated whether each course should (to some extent) cover all three perspectives or if the whole program should maintain an overall balance. The identification of potential imbalances in the program led to reflections on possible future adjustments. Constructive suggestions for course improvements emerged: lecturers acknowledged that while some highly technical programs may not require explicit socio-cultural content, integrating social implications and ethical considerations could deepen course relevance. Proposals included embedding socio-cultural expertise into existing courses or creating standalone modules, ensuring that the program remains responsive to evolving demands in the field. Through a participatory design process (von Unger, 2024) we iteratively developed the course offerings with active involvement and feedback from educators, students, and the industry perspective, fostering collaboration and ensuring quality through diverse perspectives. 3.2 A Course Template to Support the Implementation Due to the program's cross-institutional nature, courses are delivered via an open Moodle platform. The implementation of the 4C/ID model in Moodle makes use of Fig. 1. The program focuses on the technological and application perspectives of AI hardware. interactive tools and online collaboration features. A dedicated Moodle template was developed to support course implementation, offering an overall structure and building block options for the 4C/ID components. The activity “wiki”, for example, incorporates elements corresponding to two components: learning tasks (in a more active mode) and supportive information (in a more passive mode), while the "book" activity may be used to deliver supportive information. Our plan is to adapt all project-related courses progressively to this model, while conducting continuous assessment and refinement. The template marks a milestone in integrating the 4C/ID model into learning management systems (LMS). Initially created for this specific educational program using Moodle, the template is designed for broader use as a didactic resource for educators and can easily be adapted to other LMS. Beyond the AI hardware topics discussed here, this approach supports the holistic development of action-oriented skills in various application areas. 4 EXEMPLARY COURSES We examine the 4C/ID model's implementation in two courses which are part of the program and serve as illustrative examples. They were deliberately chosen for their thematic contrast—one focusing on ethical aspects of AI, the other on practical hardware design—to showcase the breadth of the educational program as well as the broad applicability of the 4C/ID model. 4.1 "Accelerating Convolutional Neural Networks Using Programmable Logic" This elective course (Fig. 1, no. 6), available to both undergraduate and graduate students, teaches students the technical concepts and skills necessary to accelerate an algorithm using a Field Programmable Gate Array (FPGA) and integrating it into a software platform. In addition to the technical aspects, the use of Convolutional Neural Networks (CNNs) offers students insights into real-world applications. The course therefore covers the technical and application perspectives as described by the Dagstuhl triangle. The course design incorporates hybrid weekly lectures, inperson help sessions, online support materials, and guided programming videos, aligning with the 4C/ID components as follows: Learning tasks: The course is organized into learning tasks: students begin by implementing small, independent kernels, and as the course progresses, the level of assistance is gradually reduced, culminating in students designing their own FPGA algorithm. Supportive information: Algorithm acceleration strategies and FPGA capabilities are covered in lectures and illustrated by an explanation of the thought process (cognitive strategies) in guided programming videos. Procedural information: Non-routine skills for programming FPGAs, such as syntax and compiler use, are provided via exercise sheets and reinforced in programming videos released just-in-time. Part-task practice: Initially expecting students to develop recurrent skills adequately, we identified a need for part-task practice in future course iterations after noting student challenges with repetitive FPGA tasks. The course was originally designed as an advanced course, targeting students with a strong computer engineering foundation. However, the insight that many students lacked essential knowledge led to a redesign to include prerequisite skills. This highlights the need for a comprehensive AI hardware-focused CS curriculum that addresses diverse student requirements. 4.2 "Ethics for Nerds" Building on prior local seminars, a cross-institutional seminar on AI and ethics (Fig. 1, no. 8) was developed as an elective course for undergraduate and graduate students, focusing on topics from AI ethics in combination with scientific writing. As positioned in the Dagstuhl triangle, the seminar focuses on the application and societal-cultural prespectives of IT systems. Student-suggested topics include societal impacts of social media, AI in medical diagnostics, whistleblowing in IT, and ethical considerations in autonomous driving. The implementation employs a blended learning approach, combining asynchronous self-directed online learning with in-person block seminars that serve as both the starting point for the course and the platform for presenting results. Online elements focus on individual work within supervised tandems. The course is implemented in Moodle using the 4C/ID model as follows: Learning tasks: Students engage in scientific writing activities related to AI ethics, such as creating a literature list, presentation slides, and a written article on their selected topic. Tasks progress from stronlgy guided activities, such as literature research, to working independently on a self-selected topic. Supportive information: Foundational knowledge and concepts are delivered via presentation slides, video recordings, and topic-specific publications, which students are required to research on their own within the framework of the learning tasks. Furthermore, regular peer assessments are conducted where students assess each other's results from the learning tasks while at the same time comparing their own results to their peers'. This feedback serves as supportive information for subsequent learning tasks. Procedural information: Tooltips and checklists help ensure compliance with submission requirements. Exercises with automatic feedback assist students in acquiring the concepts addressed by the supportive information. Part-task practice: Optional reinforcement tasks at challenging stages include assessments with automatic evaluation or peer feedback. The methodology was adapted to changing teaching conditions, emphasizing selfdirected learning in a cross-institutional setting. Over the years, we have adjusted the allocation of learning tasks to the face-to-face and online parts of the course in order to make the most of the valuable face-to-face sessions. 5 EVALUATION OF THE COURSES We optimized the didactic design of the courses using a multifaceted approach consisting of participatory workshops (as described in section 3.1) and collegial feedback in the design phase, and regular teaching evaluations in the implementation phase. Combined, these methods enabled a comprehensive and iterative refinement process, as described below. 5.1 Collegial Feedback Esteemed senior researchers in computer engineering and media-based instructional design provided valuable feedback. This collegial feedback involved an assessment of the 4C/ID model’s appropriateness and effectiveness and a careful examination of the courses' relevance, structure, and content. Positive feedback from involved experts highlighted the courses' clarity and relevance and praised the 4C/ID model for its hands-on approach. Critical observations identified gaps and provided constructive suggestions for enhancing content, teaching methods, and assessment strategies. This feedback informed targeted adjustments, reinforcing the courses' contemporary relevance and recommending development steps to adapt to evolving requirements. This iterative process established a strong foundation for the ongoing quality assurance of the courses and the didactic approach. 5.2 Teaching Evaluation Results As a second pillar, insights from regular teaching evaluations conducted at the participating universities were used to establish a continuous feedback loop (Arnold, 2000). The courses received generally positive feedback from students. In the evaluation of the course "Accelerating Convolutional Neural Networks Using Programmable Logic," 4 of 13 participants responded, finding the course challenging and engaging and valuing in-person feedback. However, they criticized the workload as excessive, and highlighted issues with time management, noting too much time spent on simpler early tasks and insufficient time for later, more complex ones. In the following iteration, the workload was reduced and time allocation shifted between modules. 8 out of 15 students provided feedback, with six students rating the course as difficult—on a scale of very easy to very difficult—and two as spot-on, showing a clear improvement.In the course "Ethics for Nerds," on-site surveys with 23 resp. 20 students were conducted in the form of flash interviews with all participants at the second block seminar near the end of the semester. The interviews revealed minimal criticism; the overall assessment of the topic and structure was very positive. Students appreciated the tandem interactions and the cross-institutional collaboration, expressing the wish for more opportunities like these. In addition, a Teaching Analysis Poll (TAP) (Johannsen & Meyer, 2023) was conducted during the second block seminar in the last iteration of the course, where students discussed key aspects of the course in small, unsupervised groups. Responses were reviewed in a plenary session, where students prioritized suggestions for change. The TAP evaluation, combined with instructor reflections, identified three areas for future improvement (Gursch et al., 2025): 1. What practices should be maintained? Students appreciated in-depth discussions, individual topic preparation, cultural exchange, and learning from others' presentations. 2. What aspects of teaching can be improved? Suggestions included avoiding seminars close to exams, clearer submission guidelines, enhanced supervisor meetings, improved peer feedback instructions, and better Moodle usability (simplified course navigation). 3. Where do differing opinions exist? Opinions on supervision varied, with some students satisfied and others calling for more guidance. Similarly, while many praised the block seminar format for its focus and intensity, some found it overly demanding. Among the 43 participants enrolled from both universities, nearly half (47%) earned an excellent grade, while over a third (39%) achieved good or very good grades, marking a performance significantly superior to that of comparable courses with similar learning objectives. The course had a drop-out rate of 13% (4 students), primarily due to personal factors, such as illness or significant family events, unrelated to course content or design. As of the submission date, evaluations from four additional instructors covering a total of ten courses among the offerings, involving 50 students, have been conducted alongside our own evaluations. Across these courses, students predominantly commended the quality of guidance, the practical exercises, and the engaging topics, while noting improvements needed in organizational aspects like timing and coordination of assignments, as well as the overlap between project work and examination periods. We will continue to apply evaluation methods to enhance the quality of the offerings. 5.3 Discussion The positive evaluation results affirm the value of participatory design and collegial feedback, with the method's effectiveness validated by good to excellent learning outcomes across individual courses. Students responded positively to online collaborative work methodologies, such as peer feedback. This suggests that engaging students in shared tasks can enhance the learning experience. However, issues related to Moodle usability were identified, indicating a need for improvements to better support the intricacies of inter-institutional collaboration. In response to this feedback, we restructured the Moodle template to make it more clear and accessible. In addition, organizational challenges such as the timing and coordination of assignments, and the overlap between project work and examination periods, were highlighted as areas that need attention. Current iterations of the courses have been adjusted accordingly. In addition to student feedback, the lecturers also received the didactic 4C/ID model positively, recognizing its structured approach and applicability to real-world tasks from various disciplines. These findings underscore the potential of media-based instructional design to enrich teaching practices and enable cross-institutional collaboration while recognizing the necessity to tackle technical and organizational barriers for broader and more effective implementation. Recognizing the inherent limitations of the methods used is essential for a comprehensive interpretation of the results. Participatory design, while valuable, can present challenges such as coordination difficulties and time constraints, which may limit the depth of stakeholder engagement. Similarly, collegial feedback is inherently subjective, requiring careful analysis to ensure reliability. Teaching evaluations do not yet encompass the full spectrum of the courses. Moreover, relying primarily on student feedback to assess success may result in overlooking other critical indicators, such as long-term career outcomes or the development of interdisciplinary skills. These aspects should be monitored in the future to provide a more holistic evaluation of the program. 6 CONCLUSIONS AND IMPLICATIONS The demand for AI hardware courses is driven by AI's transformative impact on digitization. We address this gap by integrating courses that cover essential competencies in AI, computer engineering, and selected application fields into a cross-institutional blended-learning program. In alignment with the Dagstuhl triangle, socio-cultural aspects are incorporated alongside technical and application aspects. The implementation of the 4C/ID model promotes the development of complex skills and supports individualized learning paths. Two courses were presented in more depth as examples, focusing on technical skills in FPGA acceleration and ethical considerations in AI, respectively. Ongoing refinement ensures the continued evolution of the courses. Future work will involve further evaluations of the courses, expanding the program to other universities, potentially developing it into a full graduate program and contributing to current curriculum recommendations to enhance AI hardware education.